Cerebras CS-4 skips training entirely, and that is the interesting part
750 petaflops and 7.2 terabits per second of I/O, aimed at inference rather than training. The specification is a thesis about where the money is going.
Cerebras announced the CS-4, a rack-scale accelerator rated at 750 petaflops with 7.2 terabits per second of I/O, available in Q3 2026. The specification that matters is not either number. It is the positioning: this system is for inference, explicitly not training.
Why that is a real decision and not marketing
Training and inference stress a machine in almost opposite ways, and a design can only be excellent at one.
| Training | Inference | |
|---|---|---|
| Dominant cost | Sustained floating-point throughput | Memory bandwidth and latency per token |
| Batching | Large batches, latency irrelevant | Small or single batches, latency is the product |
| Job shape | Weeks-long, one job, checkpointed | Milliseconds, millions of independent jobs |
| Failure tolerance | Restart from checkpoint | A dropped request is a user-visible error |
| Scaling axis | More FLOPs across more nodes | More concurrent streams at fixed latency |
Autoregressive decoding is the crux. Generating each token requires reading the model weights again. At batch size one, a GPU spends most of its time waiting on memory while its arithmetic units idle — you are not compute-bound, you are bandwidth-bound, and buying more FLOPs buys nothing. This is why inference economics are so counterintuitive: the expensive part is moving weights, not multiplying them.
Cerebras’ wafer-scale approach attacks exactly that. Keeping weights in enormous on-wafer SRAM rather than streaming them from external HBM removes the trip that dominates decode latency. The 7.2 Tb/s I/O figure is the other half — feeding a machine that fast requires not starving it at the edges.
The market timing is the argument
Training demand is concentrated in a handful of labs, lumpy, and increasingly financed by the arrangements that made Nvidia’s credit guarantee newsworthy. Inference demand is diffuse, recurring, and grows with every deployed product — it looks like utility load rather than capital projects.
If you are not going to beat Nvidia on training — and nobody has — then building for the workload that is becoming a metered service rather than a capex event is the defensible half of the market. It also happens to be the half that Stripe just paid $7 billion to sit in front of.
Training is a project. Inference is a bill. Businesses built on bills are worth more.
What the spec sheet does not tell you
Three caveats before anyone reorganises a procurement plan:
- Petaflops at what precision? A 750 PFLOP figure means very different things at FP16, FP8 or FP4. Inference increasingly runs quantised, so headline numbers are usually quoted at the most favourable format. Ask.
- Tokens per second per dollar is the only metric that matters. Rack-scale systems win on aggregate throughput and can still lose on cost per token against dense commodity GPU fleets.
- Software is the moat, and it is not Cerebras’. CUDA’s advantage is a decade of kernels, and every serving framework is tuned for it first. Novel architectures historically die here, not on silicon.
Q3 2026 availability means real benchmarks are close. Judge it on independent tokens-per-second-per-dollar at a fixed latency target, and ignore everything else.